昇思25天学习打卡营第3天
文章目录
模型训练
模型训练一般分为四个步骤:
- 构建数据集。
- 定义神经网络模型。
- 定义超参、损失函数及优化器。
- 输入数据集进行训练与评估。
现在我们有了数据集和模型后,可以进行模型的训练与评估。
构建数据集
首先从数据集 Dataset加载代码,构建数据集。
%%capture captured_output
# 实验环境已经预装了mindspore==2.2.14,如需更换mindspore版本,可更改下面mindspore的版本号
!pip uninstall mindspore -y
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14
import mindspore
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset
# Download data from open datasets
from download import download
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/" \
"notebook/datasets/MNIST_Data.zip"
path = download(url, "./", kind="zip", replace=True)
def datapipe(path, batch_size):
image_transforms = [
vision.Rescale(1.0 / 255.0, 0),
vision.Normalize(mean=(0.1307,), std=(0.3081,)),
vision.HWC2CHW()
]
label_transform = transforms.TypeCast(mindspore.int32)
dataset = MnistDataset(path)
dataset = dataset.map(image_transforms, 'image')
dataset = dataset.map(label_transform, 'label')
dataset = dataset.batch(batch_size)
return dataset
train_dataset = datapipe('MNIST_Data/train', batch_size=64)
test_dataset = datapipe('MNIST_Data/test', batch_size=64)
定义神经网络模型
从网络构建中加载代码,构建一个神经网络模型。
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
model = Network()
定义超参、损失函数和优化器
超参
超参(Hyperparameters)是可以调整的参数,可以控制模型训练优化的过程,不同的超参数值可能会影响模型训练和收敛速度。目前深度学习模型多采用批量随机梯度下降算法进行优化,随机梯度下降算法的原理如下:
w t + 1 = w t − η 1 n ∑ x ∈ B ∇ l ( x , w t ) w_{t+1}=w_{t}-\eta \frac{1}{n} \sum_{x \in \mathcal{B}} \nabla l\left(x, w_{t}\right) wt+1=wt−ηn1x∈B∑∇l(x,wt)
公式中, n n n是批量大小(batch size), η η η是学习率(learning rate)。另外, w t w_{t} wt为训练轮次 t t t中的权重参数, ∇ l \nabla l ∇l为损失函数的导数。除了梯度本身,这两个因子直接决定了模型的权重更新,从优化本身来看,它们是影响模型性能收敛最重要的参数。一般会定义以下超参用于训练:
-
训练轮次(epoch):训练时遍历数据集的次数。
-
批次大小(batch size):数据集进行分批读取训练,设定每个批次数据的大小。batch size过小,花费时间多,同时梯度震荡严重,不利于收敛;batch size过大,不同batch的梯度方向没有任何变化,容易陷入局部极小值,因此需要选择合适的batch size,可以有效提高模型精度、全局收敛。
-
学习率(learning rate):如果学习率偏小,会导致收敛的速度变慢,如果学习率偏大,则可能会导致训练不收敛等不可预测的结果。梯度下降法被广泛应用在最小化模型误差的参数优化算法上。梯度下降法通过多次迭代,并在每一步中最小化损失函数来预估模型的参数。学习率就是在迭代过程中,会控制模型的学习进度。
epochs = 3
batch_size = 64
learning_rate = 1e-2
损失函数
损失函数(loss function)用于评估模型的预测值(logits)和目标值(targets)之间的误差。训练模型时,随机初始化的神经网络模型开始时会预测出错误的结果。损失函数会评估预测结果与目标值的相异程度,模型训练的目标即为降低损失函数求得的误差。
常见的损失函数包括用于回归任务的nn.MSELoss
(均方误差)和用于分类的nn.NLLLoss
(负对数似然)等。 nn.CrossEntropyLoss
结合了nn.LogSoftmax
和nn.NLLLoss
,可以对logits 进行归一化并计算预测误差。
loss_fn = nn.CrossEntropyLoss()
优化器
模型优化(Optimization)是在每个训练步骤中调整模型参数以减少模型误差的过程。MindSpore提供多种优化算法的实现,称之为优化器(Optimizer)。优化器内部定义了模型的参数优化过程(即梯度如何更新至模型参数),所有优化逻辑都封装在优化器对象中。在这里,我们使用SGD(Stochastic Gradient Descent)优化器。
我们通过model.trainable_params()
方法获得模型的可训练参数,并传入学习率超参来初始化优化器。
optimizer = nn.SGD(model.trainable_params(), learning_rate=learning_rate)
在训练过程中,通过微分函数可计算获得参数对应的梯度,将其传入优化器中即可实现参数优化,具体形态如下:
grads = grad_fn(inputs)
optimizer(grads)
训练与评估
设置了超参、损失函数和优化器后,我们就可以循环输入数据来训练模型。一次数据集的完整迭代循环称为一轮(epoch)。每轮执行训练时包括两个步骤:
- 训练:迭代训练数据集,并尝试收敛到最佳参数。
- 验证/测试:迭代测试数据集,以检查模型性能是否提升。
接下来我们定义用于训练的train_loop
函数和用于测试的test_loop
函数。
使用函数式自动微分,需先定义正向函数forward_fn
,使用value_and_grad获得微分函数grad_fn
。然后,我们将微分函数和优化器的执行封装为train_step
函数,接下来循环迭代数据集进行训练即可。
# Define forward function
def forward_fn(data, label):
logits = model(data)
loss = loss_fn(logits, label)
return loss, logits
# Get gradient function
grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
# Define function of one-step training
def train_step(data, label):
(loss, _), grads = grad_fn(data, label)
optimizer(grads)
return loss
def train_loop(model, dataset):
size = dataset.get_dataset_size()
model.set_train()
for batch, (data, label) in enumerate(dataset.create_tuple_iterator()):
loss = train_step(data, label)
if batch % 100 == 0:
loss, current = loss.asnumpy(), batch
print(f"loss: {loss:>7f} [{current:>3d}/{size:>3d}]")
test_loop
函数同样需循环遍历数据集,调用模型计算loss和Accuray并返回最终结果。
def test_loop(model, dataset, loss_fn):
num_batches = dataset.get_dataset_size()
model.set_train(False)
total, test_loss, correct = 0, 0, 0
for data, label in dataset.create_tuple_iterator():
pred = model(data)
total += len(data)
test_loss += loss_fn(pred, label).asnumpy()
correct += (pred.argmax(1) == label).asnumpy().sum()
test_loss /= num_batches
correct /= total
print(f"Test: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
我们将实例化的损失函数和优化器传入train_loop
和test_loop
中。训练3轮并输出loss和Accuracy,查看性能变化。
loss_fn = nn.CrossEntropyLoss()
optimizer = nn.SGD(model.trainable_params(), learning_rate=learning_rate)
for t in range(epochs):
print(f"Epoch {t+1}\n-------------------------------")
train_loop(model, train_dataset)
test_loop(model, test_dataset, loss_fn)
print("Done!")
结果输出:
Epoch 1
-------------------------------
loss: 2.307103 [ 0/938]
loss: 1.758749 [100/938]
loss: 0.928497 [200/938]
loss: 0.677495 [300/938]
loss: 0.463961 [400/938]
loss: 0.343024 [500/938]
loss: 0.354307 [600/938]
loss: 0.518067 [700/938]
loss: 0.232976 [800/938]
loss: 0.343765 [900/938]
Test:
Accuracy: 90.8%, Avg loss: 0.315937
Epoch 2
-------------------------------
loss: 0.286438 [ 0/938]
loss: 0.472248 [100/938]
loss: 0.312173 [200/938]
loss: 0.259874 [300/938]
loss: 0.457914 [400/938]
loss: 0.193126 [500/938]
loss: 0.384219 [600/938]
loss: 0.215655 [700/938]
loss: 0.236416 [800/938]
loss: 0.200132 [900/938]
Test:
Accuracy: 92.9%, Avg loss: 0.249953
Epoch 3
-------------------------------
loss: 0.291161 [ 0/938]
loss: 0.266009 [100/938]
loss: 0.092896 [200/938]
loss: 0.185963 [300/938]
loss: 0.194996 [400/938]
loss: 0.358232 [500/938]
loss: 0.339983 [600/938]
loss: 0.201048 [700/938]
loss: 0.192553 [800/938]
loss: 0.301402 [900/938]
Test:
Accuracy: 93.7%, Avg loss: 0.212539
Done!
保存与加载
上一章节主要介绍了如何调整超参数,并进行网络模型训练。在训练网络模型的过程中,实际上我们希望保存中间和最后的结果,用于微调(fine-tune)和后续的模型推理与部署,本章节我们将介绍如何保存与加载模型。
%%capture captured_output
# 实验环境已经预装了mindspore==2.2.14,如需更换mindspore版本,可更改下面mindspore的版本号
!pip uninstall mindspore -y
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14
import numpy as np
import mindspore
from mindspore import nn
from mindspore import Tensor
def network():
model = nn.SequentialCell(
nn.Flatten(),
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10))
return model
保存和加载模型权重
保存模型使用save_checkpoint
接口,传入网络和指定的保存路径:
model = network()
mindspore.save_checkpoint(model, "model.ckpt")
要加载模型权重,需要先创建相同模型的实例,然后使用load_checkpoint
和load_param_into_net
方法加载参数。
model = network()
param_dict = mindspore.load_checkpoint("model.ckpt")
param_not_load, _ = mindspore.load_param_into_net(model, param_dict)
print(param_not_load)
结果输出:
[]
param_not_load
是未被加载的参数列表,为空时代表所有参数均加载成功。
保存和加载MindIR
除Checkpoint外,MindSpore提供了云侧(训练)和端侧(推理)统一的中间表示(Intermediate Representation,IR)。可使用export
接口直接将模型保存为MindIR。
model = network()
inputs = Tensor(np.ones([1, 1, 28, 28]).astype(np.float32))
mindspore.export(model, inputs, file_name="model", file_format="MINDIR")
MindIR同时保存了Checkpoint和模型结构,因此需要定义输入Tensor来获取输入shape。
已有的MindIR模型可以方便地通过load
接口加载,传入nn.GraphCell
即可进行推理。
nn.GraphCell
仅支持图模式。
mindspore.set_context(mode=mindspore.GRAPH_MODE)
graph = mindspore.load("model.mindir")
model = nn.GraphCell(graph)
outputs = model(inputs)
print(outputs.shape)
结果输出:
(1, 10)
使用静态图加速
背景介绍
AI编译框架分为两种运行模式,分别是动态图模式以及静态图模式。MindSpore默认情况下是以动态图模式运行,但也支持手工切换为静态图模式。两种运行模式的详细介绍如下:
动态图模式
动态图的特点是计算图的构建和计算同时发生(Define by run),其符合Python的解释执行方式,在计算图中定义一个Tensor时,其值就已经被计算且确定,因此在调试模型时较为方便,能够实时得到中间结果的值,但由于所有节点都需要被保存,导致难以对整个计算图进行优化。
在MindSpore中,动态图模式又被称为PyNative模式。由于动态图的解释执行特性,在脚本开发和网络流程调试过程中,推荐使用动态图模式进行调试。
如需要手动控制框架采用PyNative模式,可以通过以下代码进行网络构建:
%%capture captured_output
# 实验环境已经预装了mindspore==2.2.14,如需更换mindspore版本,可更改下面mindspore的版本号
!pip uninstall mindspore -y
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
ms.set_context(mode=ms.PYNATIVE_MODE) # 使用set_context进行动态图模式的配置
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
model = Network()
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
output = model(input)
print(output)
结果输出:
[[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]
[-0.07497295 -0.01996819 0.11828961 -0.07505784 -0.06651713 0.06720132
-0.08175808 -0.21639574 -0.04644566 0.00188687]]
静态图模式
相较于动态图而言,静态图的特点是将计算图的构建和实际计算分开(Define and run)。有关静态图模式的运行原理,可以参考静态图语法支持。
在MindSpore中,静态图模式又被称为Graph模式,在Graph模式下,基于图优化、计算图整图下沉等技术,编译器可以针对图进行全局的优化,获得较好的性能,因此比较适合网络固定且需要高性能的场景。
如需要手动控制框架采用静态图模式,可以通过以下代码进行网络构建:
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
ms.set_context(mode=ms.GRAPH_MODE) # 使用set_context进行运行静态图模式的配置
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
model = Network()
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
output = model(input)
print(output)
结果输出:
[[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]
[-0.13221586 -0.1204772 -0.02582647 -0.08580489 -0.01784871 -0.00198492
0.05919297 -0.09395744 0.05590799 0.0362997 ]]
静态图模式的使用场景
MindSpore编译器重点面向Tensor数据的计算以及其微分处理。因此使用MindSpore API以及基于Tensor对象的操作更适合使用静态图编译优化。其他操作虽然可以部分入图编译,但实际优化作用有限。另外,静态图模式先编译后执行的模式导致其存在编译耗时。因此,如果函数无需反复执行,那么使用静态图加速也可能没有价值。
有关使用静态图来进行网络编译的示例,请参考网络构建。
静态图模式开启方式
通常情况下,由于动态图的灵活性,我们会选择使用PyNative模式来进行自由的神经网络构建,以实现模型的创新和优化。但是当需要进行性能加速时,我们需要对神经网络部分或整体进行加速。MindSpore提供了两种切换为图模式的方式,分别是基于装饰器的开启方式以及基于全局context的开启方式。
基于装饰器的开启方式
MindSpore提供了jit装饰器,可以通过修饰Python函数或者Python类的成员函数使其被编译成计算图,通过图优化等技术提高运行速度。此时我们可以简单的对想要进行性能优化的模块进行图编译加速,而模型其他部分,仍旧使用解释执行方式,不丢失动态图的灵活性。无论全局context是设置成静态图模式还是动态图模式,被jit修饰的部分始终会以静态图模式进行运行。
在需要对Tensor的某些运算进行编译加速时,可以在其定义的函数上使用jit修饰器,在调用该函数时,该模块自动被编译为静态图。需要注意的是,jit装饰器只能用来修饰函数,无法对类进行修饰。jit的使用示例如下:
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
@ms.jit # 使用ms.jit装饰器,使被装饰的函数以静态图模式运行
def run(x):
model = Network()
return model(x)
output = run(input)
print(output)
除使用修饰器外,也可使用函数变换方式调用jit方法,示例如下:
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
def run(x):
model = Network()
return model(x)
run_with_jit = ms.jit(run) # 通过调用jit将函数转换为以静态图方式执行
output = run(input)
print(output)
当我们需要对神经网络的某部分进行加速时,可以直接在construct方法上使用jit修饰器,在调用实例化对象时,该模块自动被编译为静态图。示例如下:
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
@ms.jit # 使用ms.jit装饰器,使被装饰的函数以静态图模式运行
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
model = Network()
output = model(input)
print(output)
基于context的开启方式
context模式是一种全局的设置模式。代码示例如下:
import numpy as np
import mindspore as ms
from mindspore import nn, Tensor
ms.set_context(mode=ms.GRAPH_MODE) # 使用set_context进行运行静态图模式的配置
class Network(nn.Cell):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.dense_relu_sequential = nn.SequentialCell(
nn.Dense(28*28, 512),
nn.ReLU(),
nn.Dense(512, 512),
nn.ReLU(),
nn.Dense(512, 10)
)
def construct(self, x):
x = self.flatten(x)
logits = self.dense_relu_sequential(x)
return logits
model = Network()
input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))
output = model(input)
print(output)
静态图的语法约束
在Graph模式下,Python代码并不是由Python解释器去执行,而是将代码编译成静态计算图,然后执行静态计算图。因此,编译器无法支持全量的Python语法。MindSpore的静态图编译器维护了Python常用语法子集,以支持神经网络的构建及训练。详情可参考静态图语法支持。
JitConfig配置选项
在图模式下,可以通过使用JitConfig配置选项来一定程度的自定义编译流程,目前JitConfig支持的配置参数如下:
- jit_level: 用于控制优化等级。
- exec_mode: 用于控制模型执行方式。
- jit_syntax_level: 设置静态图语法支持级别,详细介绍请见静态图语法支持。
静态图高级编程技巧
使用静态图高级编程技巧可以有效地提高编译效率以及执行效率,并可以使程序运行的更加稳定。详情可参考静态图高级编程技巧。
打卡记录